Lightweight oral image AI diagnosis system

Abstract

The invention relates to the technical field of oral medical image processing, in particular to a lightweight oral image AI diagnosis system, which comprises front-end acquisition equipment and a server, the front-end acquisition equipment is used for acquiring an oral cavity image to be diagnosed, and the server comprises a data storage module, a model training module, a self-adaptive reasoning module and a result output module; the data storage module is used for storing historical oral images and corresponding diagnosis results; the model training module is used for constructing a teacher model and student model architecture, and enabling a lightweight student model to learn the feature extraction capability of a high-precision teacher model through a distillation loss function; the self-adaptive reasoning module is used for dynamically adjusting a reasoning strategy according to the difficulty level of the image; and the result output module is used for generating a clinical diagnosis report including tooth position identification and focus types and positions. According to the system, the balance of light weight and high precision is realized, the reasoning efficiency is improved, the clinical application threshold is reduced, and the model diagnosis precision can be continuously optimized through historical data.

Description

Technical Field

[0001] This invention relates to the field of oral medical image processing technology, specifically to a lightweight oral image AI diagnostic system. Background Technology

[0002] AI-assisted diagnostic technology for oral imaging is an important research direction at the intersection of clinical medicine and artificial intelligence. Its core is to automate diagnostic tasks such as tooth position recognition and lesion detection through deep learning models. For example, Chinese invention patent CN118365964B discloses a method for identifying tooth position and periapical inflammation based on deep learning in panoramic radiographs. This method uses a deep learning network to extract image features of teeth and periapical inflammation, and combines this with a bidirectional long short-term memory network (BiLSTM) for sequence modeling, achieving joint recognition of tooth position numbers and periapical inflammation. Another example is Chinese invention patent CN119273632B, which discloses an oral disease detection method based on a large-scale artificial intelligence model. This method utilizes a large-scale pre-trained deep learning model (such as the YOLOv8 network) to extract features from oral images, and integrates image and clinical text information through multimodal fusion technology, significantly improving the accuracy of identifying complex lesions, especially in the detection of rare diseases and complications.

[0003] However, existing AI diagnostic technologies for oral imaging still have the following shortcomings in practical applications: Existing high-precision models such as Mask R-CNN and YOLOv8 networks have a large number of parameters, and inference from a single image requires a lot of computational resources; at the same time, most existing technologies perform complete feature extraction and sequence modeling processes on all images, which wastes computational power for simple images without lesions. Furthermore, in oral images, even if there are lesions, the teeth and lesion areas usually only account for 30%-40% of the entire image, with the rest being irrelevant areas such as the gums and background. Performing complete convolution and feature extraction operations on the entire image increases the proportion of invalid computation, with most computational resources being used for background region processing, reducing the model's processing efficiency; in addition, the output results of existing technologies are mostly feature vectors and classification probabilities. Although some existing technologies provide bounding box annotations, they do not integrate tooth position numbers and clinical diagnostic terminology. Primary care physicians need to manually match tooth positions with lesions, which increases the threshold for clinical application and reduces the efficiency of use.

[0004] In summary, existing technologies struggle to achieve lightweight models while maintaining diagnostic accuracy, and the output format does not meet clinical needs, hindering the widespread adoption of intelligent diagnostic technologies in resource-constrained scenarios. Therefore, there is an urgent need for an oral imaging diagnostic system that balances lightweight design, high efficiency, high accuracy, and standardized output. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a lightweight oral imaging AI diagnostic system that enables lightweight and efficient reasoning and generates clinically usable diagnostic reports.

[0006] The basic solution provided by this invention is a lightweight oral imaging AI diagnostic system, comprising a front-end acquisition device and a server. The front-end acquisition device is used to acquire oral images to be diagnosed, and the server includes: The data storage module is used to store historical oral images and corresponding diagnostic results; The model training module is used to construct the teacher model and student model architecture. The teacher model is built based on the YOLOv8 network and trained using historical oral images and corresponding diagnostic results. The student model is built based on a lightweight convolutional network and trained to learn the feature extraction ability of the teacher model through a distillation loss function. The number of parameters of the student model is smaller than that of the teacher model. An adaptive inference module includes a difficulty perception unit and a dynamic adjustment unit. The difficulty perception unit is used to evaluate the difficulty level of the oral image, which includes easy, medium and complex levels. The dynamic adjustment unit is used to adjust the inference strategy of the student model according to the difficulty level. The results output module is used to receive the prediction results of the student model, generate and output a diagnostic report including tooth position recognition results, lesion type and location.

[0007] The principle of this invention is as follows: Oral images (including common types such as panoramic radiographs and periapical radiographs) to be diagnosed are acquired through front-end acquisition devices (such as digital dental X-ray machines, CBCT equipment, etc.) and sent to a server. The server's data storage module receives and stores newly acquired images to be diagnosed, and also permanently stores historical oral images and corresponding clinical diagnostic results (including tooth position annotations, lesion types, doctor's diagnostic conclusions, etc.), forming a structured medical database. This database not only provides high-quality labeled samples for model training but also supports retrospective comparison and diagnostic consistency verification of subsequent cases, laying a data foundation for continuous system optimization. In the model training module of the server, teacher and student model architectures are constructed. The teacher model is built based on the YOLOv8 network and is fully trained on historical image datasets with the help of its powerful feature extraction capabilities to form a high-precision recognition capability for tooth position and lesions in oral images. The student model is built based on a lightweight convolutional network (such as a depthwise separable convolutional architecture) and learns the feature extraction capability of the teacher model through a distillation loss function. Thus, with a parameter count much smaller than that of the teacher model (usually 10%-20% of that of the teacher model), it retains diagnostic accuracy close to that of the teacher model. In the adaptive inference module, a difficulty-aware unit performs multi-dimensional evaluation of the input oral images: by analyzing image clarity, occlusion level, and lesion complexity, the images are categorized into three difficulty levels: simple, medium, and complex. Subsequently, a dynamic adjustment unit adapts the student model's inference strategy according to the difficulty level: for simple images with clear features and minimal interference, a simplified model is used to improve speed; for medium-quality images, the full model is used for inference to ensure basic accuracy; for blurry, highly occluded, or complex lesion images, the output of the teacher model is introduced to correct the inference based on the full model, ensuring diagnostic reliability. This tiered inference strategy avoids wasting computational resources on simple images while ensuring diagnostic quality for complex images. The final output module transforms the student model's predictions (such as tooth position coordinates, lesion type, confidence level, and other raw data) into a clinically readable diagnostic report. By integrating image visualization annotations (such as the correspondence between tooth position numbers and lesion areas) with structured text descriptions (such as the health status of each tooth, lesion type, and confidence level), it directly outputs a report format that conforms to clinical diagnosis and treatment habits, lowering the threshold for clinical application.

[0008] The beneficial effects of this invention are as follows: Through the architecture of teacher and student models, the student model, with a significantly reduced number of parameters (far smaller than the teacher model), can still learn the feature extraction capabilities of the teacher model through distillation, maintaining high accuracy in inference results while balancing lightweight design, high efficiency, and high precision. The adaptive inference module adjusts strategies based on image difficulty, avoiding wasted computational power on simple images and enhancing processing for complex images, thus improving the system's processing efficiency in diverse scenarios. The result output module directly generates standardized reports including tooth position and lesion information, lowering the barrier to entry for primary care physicians and addressing the disconnect between existing technology outputs and clinical needs. Furthermore, the historical data accumulated by the data storage module can be used for model iteration and optimization. As the data volume increases, the performance of both the teacher and student models can be continuously optimized through retraining, enabling the system's diagnostic accuracy to continuously improve with clinical application and possessing long-term clinical value.

[0009] Furthermore, the teacher model adopts the YOLOv8-Large architecture, and the student model is based on the MobileNetV3 architecture with the addition of an oral cavity feature attention module; the distillation loss function includes classification loss, localization loss and feature matching loss, the classification loss adopts the cross-entropy loss function, the localization loss adopts the CIOU loss function, and the feature matching loss adopts the L2 distance loss of the intermediate layer features of the teacher model and the student model.

[0010] Furthermore, the difficulty level of the oral image is determined based on a combination of the image's clarity level, occlusion rate level, and lesion complexity level. The clarity level includes clear, normal, and blurry; the occlusion rate level includes low occlusion, normal occlusion, and high occlusion; and the lesion complexity level includes simple, moderate, and complex. Determining the image difficulty level through a combination of these three dimensions—clarity, occlusion rate, and lesion complexity—covers the core elements of image quality (clarity), interfering factors (occlusion rate), and diagnostic complexity (lesion complexity). This avoids the one-sidedness of single-dimensional assessment, makes the difficulty classification more objective, provides an accurate basis for adaptive reasoning, and reduces strategy misjudgment.

[0011] Furthermore, the clarity level is obtained in the following way: the oral image is grayscaled, the Sobel operator is used to calculate the gradient values ​​in the horizontal and vertical directions, the square root of the sum of the squares of the gradient values ​​is taken as the edge gradient value, the mean of the edge gradient values ​​of the whole image is calculated, and the current oral image clarity level is output according to the calculated mean and the preset mean and clarity level correspondence rules. The occlusion rate is obtained by converting the oral cavity image into a grayscale image. Areas with grayscale values ​​exceeding a preset value are designated as bright areas. The percentage of pixels in the bright areas is counted relative to the total number of pixels in the entire image. Based on the calculated percentage and the preset rule for corresponding percentage and occlusion rate level, the current oral cavity image occlusion rate level is output. The lesion complexity is obtained in the following way: the contour of the lesion region in the oral image is extracted using the Canny edge detection algorithm, the ratio of the perimeter of the contour to the area of ​​the region enclosed by the contour is calculated, and the lesion complexity level of the current oral image is output according to the calculated ratio and the preset ratio and lesion complexity level correspondence rules.

[0012] In this scheme, sharpness is quantified using the mean gradient of the Sobel operator to avoid subjective judgment errors; occlusion rate is quantified by the proportion of bright areas to accurately reflect the occlusion situation; and lesion complexity is quantified by the ratio of contour perimeter to area to objectively distinguish the regularity and complexity of lesions. These quantitative standards transform difficulty assessment from qualitative to quantitative, ensuring consistency in assessment across different images and improving system stability.

[0013] Furthermore, the rules for determining the difficulty level of the oral image based on the combination of its clarity level, occlusion rate level, and lesion complexity level include: When the clarity level of the oral image is clear, the occlusion level is low, and the lesion complexity level is simple, the difficulty level of the oral image is determined to be simple. When the clarity level of the oral image is blurry, the occlusion level is high occlusion, and the lesion complexity level is complex, the difficulty level of the oral image is determined to be complex. The difficulty level of the oral imaging was determined to be moderate based on the combination of other clarity levels, occlusion levels, and lesion complexity levels.

[0014] In this scheme, the combination rules of difficulty levels make the difficulty division clear and operable, ensuring that the dynamic adjustment unit can accurately match the reasoning strategy and improve the engineering practicality of the system.

[0015] Furthermore, the reasoning strategy of the dynamic adjustment unit includes: When the difficulty level is easy, only some channels in the student model are used for inference, and the feature dimensions of each channel are adjusted to the fixed dimensions required by the output layer through the subsequent feature processing layer, so that the structure and dimensions of the inference output are consistent with those of the student model when all channels are used. When the difficulty level is medium, all channels of the student model are used for inference. When the difficulty level is complex, all channels of the student model are used for inference, and the prediction results of the student model are corrected using the output of the teacher model.

[0016] This solution employs a tiered three-level inference strategy to achieve refined allocation of computing power: for easy difficulty, some channels are used to reduce computational load; for medium difficulty, full channels are used to balance accuracy and efficiency; and for complex difficulty, full channels are used, with accuracy improved through teacher model correction. This three-level inference strategy enables the system to achieve optimal cost-effectiveness in standard images (medium difficulty) while also considering the needs of both simple and complex scenarios, further optimizing resource utilization.

[0017] Furthermore, when the difficulty level is "easy," the selected channels are determined through a selection process. The selection rules include the gradient magnitude or attention weight of each channel for feature extraction from easy-difficulty images. The feature processing layer includes a pointwise convolutional layer, used to map the output of some channels to an output dimension consistent with that of all channels during inference. The dynamic adjustment unit maintains the size of the intermediate feature map using zero-padding, ensuring compatibility of the input format for subsequent layers. In easy difficulty, channels are selected for use based on gradient magnitude or attention weight to ensure that the retained channels focus on the core features of easy images (such as the shadows of typical cavities and regular boundaries), enabling accurate diagnosis. By mapping the output dimension through pointwise convolutional layers and maintaining the feature map size with zero padding, the output structure of some channels during inference is ensured to be consistent with that of all channels.

[0018] Furthermore, the diagnostic report generated and output by the result output module includes: The visualization results of oral images are labeled with tooth position numbers, which correspond to tooth position identification results. The teeth and lesion areas are marked with masks to show the location of the lesions. The textual description includes the health status of each tooth, the type of lesion, and the corresponding confidence level.

[0019] In this solution, the diagnostic report integrates information through mask marking and text description: the mask accurately marks the teeth and lesion areas, the tooth position number intuitively corresponds to the location, and the text description includes the health status, lesion type and confidence level, eliminating the need to manually match tooth positions and lesions, thus improving clinical work efficiency. Attached Figure Description

[0020] Figure 1 This is a system module diagram of an embodiment of the lightweight oral imaging AI diagnostic system of the present invention. Detailed Implementation

[0021] The following detailed description illustrates the specific implementation method: The basic implementation examples are as follows: Figure 1 As shown: A lightweight oral imaging AI diagnostic system includes a front-end acquisition device and a server. The front-end acquisition device is used to acquire oral images to be diagnosed, and the server includes: The data storage module is used to store historical oral images and corresponding diagnostic results; The model training module is used to construct the teacher model and student model architecture. The teacher model is built based on the YOLOv8 network and trained using historical oral images and corresponding diagnostic results. The student model is built based on a lightweight convolutional network and trained to learn the feature extraction ability of the teacher model through a distillation loss function. The number of parameters of the student model is smaller than that of the teacher model. An adaptive inference module includes a difficulty perception unit and a dynamic adjustment unit. The difficulty perception unit is used to evaluate the difficulty level of the oral image, which includes easy, medium and complex levels. The dynamic adjustment unit is used to adjust the inference strategy of the student model according to the difficulty level. The results output module is used to receive the prediction results of the student model, generate and output a diagnostic report including tooth position recognition results, lesion type and location.

[0022] In this embodiment, the front-end acquisition device is an oral digital imaging device (such as a CBCT machine or an oral scanner), and the images are transmitted to the data storage module of the server via a wireless network.

[0023] The data storage module receives and stores newly acquired images to be diagnosed, and also preserves historical oral images and corresponding clinical diagnostic results (including tooth position markings, lesion types, doctor's diagnostic conclusions, etc.) for a long time, forming a structured medical database. The database not only provides high-quality labeled samples for model training, but also supports retrospective comparison of subsequent cases and verification of diagnostic consistency, laying a data foundation for continuous system optimization.

[0024] In the model training module, the teacher model training uses the YOLOv8-Large architecture, with CSPDarknet-53 as the backbone and PAN-FPN as the neck layer. The output layer predicts bounding boxes (x, y, w, h), class probabilities (16 classes in total, including 15 lesion classes and healthy individuals), and confidence scores. Training data consists of 10,000 labeled oral images (including common lesions such as caries and pulpitis), divided into training and validation sets in an 8:2 ratio. Training parameters include an initial learning rate of 0.01, a batch size of 32, 150 epochs, and the use of the SGD optimizer. Training stops when the validation set accuracy exceeds 92%. The student model training uses the MobileNetV3-Small architecture, removing the last two convolutional blocks to reduce the number of parameters (the final number of parameters is approximately 1 / 10 of the teacher model). A new oral feature attention module is added (channel attention focuses on tooth texture, and spatial attention focuses on lesion edges). Distillation training uses the following loss function:

[0025] in Cross-entropy loss (constrained category prediction). CIOU loss (constrained bounding box localization). The L2 distance loss (constrained feature distribution) is used for the teacher-student model; training parameters: learning rate 0.001, batch size 32, 150 epochs, using the Adam optimizer.

[0026] The adaptive inference module includes a difficulty perception unit and a dynamic adjustment unit. The difficulty perception unit is used to evaluate the difficulty level of the oral image. The difficulty level of the oral image includes simple, medium and complex. The difficulty level of the oral image is determined based on a combination of the clarity level, occlusion rate level and lesion complexity level of the oral image. The clarity level includes clear, normal and blurry. The occlusion rate level includes low occlusion, normal occlusion and high occlusion. The lesion complexity level includes simple, medium and complex.

[0027] The sharpness level is calculated as follows: After converting the image to grayscale, the Sobel operator is used to calculate the horizontal gradient. and vertical gradient Edge gradient value Calculate the gradient values ​​at the edges of the entire image. mean Preset rules: The resolution level is clear. The resolution level is normal. The clarity level is blurry.

[0028] The occlusion rate level is calculated as follows: The oral image is converted into a grayscale image. Areas with a grayscale value greater than or equal to 200 in the grayscale image are defined as highlighted areas (occlusion by metal restorations, etc.); the proportion of highlighted areas is calculated. Preset rules: The occlusion level is low. The occlusion level is normal occlusion. The occlusion level is high occlusion.

[0029] The complexity level of the lesion is calculated as follows: Canny edge detection (threshold [50, 150]) is used to extract the lesion contour, and the ratio of the contour perimeter P to the area A is calculated. Preset rules: The lesion complexity level was simple. The lesion complexity level was complex. The complexity level of the lesions was moderate.

[0030] The rules for determining the difficulty level of an oral image based on the combination of its clarity level, occlusion rate level, and lesion complexity level are as follows: when the clarity level of the oral image is clear, the occlusion rate level is low, and the lesion complexity level is simple, the difficulty level of the oral image is determined to be simple; when the clarity level of the oral image is blurry, the occlusion rate level is high, and the lesion complexity level is complex, the difficulty level of the oral image is determined to be complex; for other combinations of clarity level, occlusion rate level, and lesion complexity level, the difficulty level of the oral image is determined to be moderate.

[0031] The dynamic adjustment unit is used to adjust the reasoning strategy of the student model according to the difficulty level, and the reasoning strategy includes: When the difficulty level is "easy," only a subset of channels in the student model are used for inference. Subsequent feature processing layers adjust the feature dimensions of each enabled channel to the fixed dimensions required by the output layer, ensuring that the structure and dimensions of the inference output are consistent with those of the student model using all channels. Specifically, in this embodiment, the top 50% of channels in the student model based on gradient magnitude (determined by channel importance ranking during training) are enabled. A 1x1 pointwise convolution maps the feature dimensions to match the dimensions of all channels, with zero padding to maintain the feature map size.

[0032] When the difficulty level is medium, all channels of the student model are used for inference. When the difficulty level is complex, all channels of the student model are used for inference, and the prediction results of the student model are corrected by using the output of the teacher model. Specifically, the class confidence of the two models are fused (correction weight ratio 7:3, teacher model output accounts for 70%, student model output accounts for 30%) to improve diagnostic accuracy.

[0033] The diagnostic report generated and output by the results output module includes: The visualization results of oral images are labeled with tooth position numbers, which correspond to tooth position identification results. The teeth and lesion areas are marked with masks to show the location of the lesions. The textual description includes the health status of each tooth, the type of lesion, and the corresponding confidence level.

[0034] The visualization results include: marking the lesion area on the original image with a red mask (50% transparency), labeling the tooth position number with yellow numbers (FDI tooth position representation) (e.g., "11" represents the right upper central incisor), and smoothing the mask edges (Gaussian blur radius 1px) to avoid jagged edges affecting observation; The text description section specifically includes: arranged in the order of tooth position (upper right, upper left, lower left, lower right), in the format of "tooth position X: healthy status (lesion type, confidence level Y%)", for example "tooth position 36: abnormal (mesiotic caries, confidence level 93%)".

[0035] The above are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A lightweight oral imaging AI diagnostic system, characterized in that, It includes a front-end acquisition device and a server, wherein the front-end acquisition device is used to acquire oral images to be diagnosed, and the server includes: The data storage module is used to store historical oral images and corresponding diagnostic results; The model training module is used to construct the teacher model and student model architecture. The teacher model is built based on the YOLOv8 network and trained using historical oral images and corresponding diagnostic results. The student model is built based on a lightweight convolutional network and trained to learn the feature extraction ability of the teacher model through a distillation loss function. The number of parameters of the student model is smaller than that of the teacher model. An adaptive inference module includes a difficulty perception unit and a dynamic adjustment unit. The difficulty perception unit is used to evaluate the difficulty level of the oral image, which includes easy, medium and complex levels. The dynamic adjustment unit is used to adjust the inference strategy of the student model according to the difficulty level. The results output module is used to receive the prediction results of the student model, generate and output a diagnostic report including tooth position recognition results, lesion type and location.

2. The lightweight oral imaging AI diagnostic system according to claim 1, characterized in that, The teacher model adopts the YOLOv8-Large architecture, and the student model is based on MobileNetV3 with the addition of an oral cavity feature attention module. The distillation loss function includes classification loss, localization loss and feature matching loss. The classification loss adopts the cross-entropy loss function, the localization loss adopts the CIOU loss function, and the feature matching loss adopts the L2 distance loss of the intermediate layer features between the teacher model and the student model.

3. The lightweight oral imaging AI diagnostic system according to claim 1, characterized in that, The difficulty level of the oral imaging is determined based on a combination of the image clarity level, occlusion rate level, and lesion complexity level. The clarity level includes clear, normal, and blurry; the occlusion rate level includes low occlusion, normal occlusion, and high occlusion; and the lesion complexity level includes simple, moderate, and complex.

4. The lightweight oral imaging AI diagnostic system according to claim 3, characterized in that, The clarity level is obtained by: performing grayscale processing on the oral image, using the Sobel operator to calculate the gradient values ​​in the horizontal and vertical directions, taking the square root of the sum of the squares of the gradient values ​​as the edge gradient value, calculating the mean of the edge gradient values ​​of the entire image, and outputting the current clarity level of the oral image based on the calculated mean and the preset rule for the correspondence between the mean and the clarity level. The occlusion rate is obtained by converting the oral cavity image into a grayscale image. Areas with grayscale values ​​exceeding a preset value are designated as bright areas. The percentage of pixels in the bright areas is counted relative to the total number of pixels in the entire image. Based on the calculated percentage and the preset rule for corresponding percentage and occlusion rate level, the current oral cavity image occlusion rate level is output. The lesion complexity is obtained in the following way: the contour of the lesion region in the oral image is extracted using the Canny edge detection algorithm, the ratio of the perimeter of the contour to the area of ​​the region enclosed by the contour is calculated, and the lesion complexity level of the current oral image is output according to the calculated ratio and the preset ratio and lesion complexity level correspondence rules.

5. The lightweight oral imaging AI diagnostic system according to claim 3, characterized in that, The rules for determining the difficulty level of an oral image based on a combination of its clarity level, occlusion level, and lesion complexity level include: When the clarity level of the oral image is clear, the occlusion level is low, and the lesion complexity level is simple, the difficulty level of the oral image is determined to be simple. When the clarity level of the oral image is blurry, the occlusion level is high occlusion, and the lesion complexity level is complex, the difficulty level of the oral image is determined to be complex. The difficulty level of the oral imaging was determined to be moderate based on the combination of other clarity levels, occlusion levels, and lesion complexity levels.

6. The lightweight oral imaging AI diagnostic system according to claim 1, characterized in that, The reasoning strategy of the dynamic adjustment unit includes: When the difficulty level is easy, only some channels in the student model are used for inference, and the feature dimensions of each channel are adjusted to the fixed dimensions required by the output layer through subsequent feature processing layers, so that the structure and dimensions of the inference output are consistent with those of the student model when all channels are used. When the difficulty level is medium, all channels of the student model are used for inference. When the difficulty level is complex, all channels of the student model are used for inference, and the prediction results of the student model are corrected using the output of the teacher model.

7. The lightweight oral imaging AI diagnostic system according to claim 6, characterized in that, When the difficulty level is easy, the selected channels are determined through the training process. The selection rules include the gradient magnitude or attention weight of each channel for the extraction of features from easy-difficulty images. The feature processing layer includes a pointwise convolutional layer, which maps the output of some channels to the same output dimension as when inferring all channels; the dynamic adjustment unit keeps the size of the intermediate feature map unchanged by zero padding, ensuring that the input format of subsequent layers is compatible.

8. The lightweight oral imaging AI diagnostic system according to claim 1, characterized in that, The diagnostic report generated and output by the result output module includes: The visualization results of oral images are labeled with tooth position numbers, which correspond to tooth position identification results. The teeth and lesion areas are marked with masks to show the location of the lesions. The textual description includes the health status of each tooth, the type of lesion, and the corresponding confidence level.

Images